Investigations on efficient adaptation algorithms

Abstract

Efficient adaptation algorithms, which are intended to improve the performances of the LMS and the RLS algorithms are introduced. It is shown that nonlinear transformations of the input and the desired signals by a softlimiter improve the convergence speed of the LMS algorithm at no cost, with a small bias in the optimal filter coefficients. Also, the new algorithm can be used to filter a-stable non-Gaussian processes for which the conventional adaptive algorithms are useless. In a second approach, a prewhitening filter is used to increase the convergence speed of the LMS algorithm. It is shown that prewhitening does not change the relation between the input and the desired signals provided that the relation is a linear one. A low order adaptive prewhitening filter can provide significant speed up in the convergence. Finally, adaptive filtering algorithms running on roughly quantized signals are proposed to decrease the number of multiplications in the LMS and the RLS algorithms. Although, they require significantly less computations their preformances are comparable to those of the conventional LMS and RLS algorithms.